Contrastive Learning + Normalizing Flows = Anomaly Detection ᓬ(•ᴗ•)ᕒ

Introduction

This project can be broken into two primary components

• Contrastive Learning: Trains an auto-encoder to compress raw CMS event data into a meaningful, compact latent space. Uses CL to cluster events of identical labels and distance dissimilar events.
• Normalizing Flows: Uses normalizing flows to detect anomalous data from standard model physics events within the compressed representation.

Files

• configs.npy Contains relevant constants associated with normalization and project hyper-parameters in a dict.
• data_preprocessing.py All functions associated with preprocessing Delphes and raw CMS data before use. Experimentally concluded zscore and max pT normalization are optimal for Delphes.
• dnn_classifier.py First iteration of classifiers that implements a simple DNN. See normalizing flow for current use.
• graphing_module.py Repository of all graphing modules: PCA, Corner, ROC, tSNE, ect
• hyperparam_search.py Uses 💗KerasTuner💗 for hyperparameter search.
• losses.py Defines custom loss functions: reconstruction, KL, and SimCLR
• make_datasets.py Creates biased-training datasets st event representations are more distributed across tt and QCD compared to Delphes. Also provides interface for converting raw cms data into interpretable representation.
• models.py Defines models as the name suggests
• nf_classifier.py Creates initial normalizing flow class and trains model using Delphes
• predict.py Given pre-trained AE, visualizes where raw CMS data lies in the latent space.
• test.py Overlord file. Used for all initial training and plotting matters.